Data audit and problem framing
We check for leakage, duplicates, class imbalance and label noise, then define splits by site, subject or time so that test data is independent. The output is a written evaluation protocol.
Service
Artificial intelligence and machine learning development
We build machine learning systems that are measured against a baseline, trained with reproducible pipelines and monitored after release.
The problem
A model that works in a notebook can still fail in use
The training data may not match what the system sees in the field. Images of the same object, patient or field plot may sit in both the training and the test set, so the reported accuracy is inflated. Labels may differ between annotators. After release, the input distribution drifts, and nobody notices until users complain.
Decision-makers face a second difficulty: it is hard to tell a sound result from an optimistic one. A single accuracy figure says little about rare classes, a new site or poor lighting. Without a baseline, nobody can say whether a deep network is doing better than a simple rule.
We treat these as engineering problems. Before training anything we audit the data, define the evaluation protocol and agree on the metric that matters to your operation. Model selection comes after that, and a simpler model wins whenever it meets the target.
What we deliver
AI & Machine Learning sub-services
Each item can be contracted separately or combined into one programme of work.
Computer vision
Classification, detection, segmentation, tracking and OCR for images from line cameras, drones, microscopes, phones and medical scanners, including visual inspection for defects.
Language models and NLP
Document classification, information extraction, semantic search and retrieval-augmented generation, each with its own evaluation set. Open-weight models can run on your servers when data must stay in-house.
Forecasting and anomaly detection
Time-series models for demand, load and sensor data, and anomaly detectors for equipment health, network traffic and transactions, with alert thresholds tuned to the cost of a false alarm.
Edge and embedded ML
Quantisation, pruning and distillation to fit a model onto GPU modules, mobile processors or microcontrollers. Latency, memory and power are measured on the target device.
MLOps and monitoring
Versioned data and models, automated training and test pipelines, a model registry, staged release and drift monitors, so that a model can be retrained and rolled back like any other software.
Technologies and methods
Tools we use and how we use them
We choose tools that your team can maintain and that carry permissive licences. The methods matter more than the brand names.
Frameworks and runtimes
PyTorch, TensorFlow, scikit-learn, XGBoost and LightGBM for training; Hugging Face Transformers for language and vision models; ONNX Runtime, TensorRT and TensorFlow Lite for deployment.
Evaluation
Grouped and time-based cross-validation, external test sets, bootstrap confidence intervals, calibration curves, per-class and per-subgroup reporting, and ablation studies that show which component earns its place.
Explainability and uncertainty
Grad-CAM and SHAP to inspect what a model relies on, and conformal prediction to produce prediction sets with a stated error rate, so the system can abstain and route a case to a person.
Pipelines
Data versioning with DVC, experiment tracking with MLflow, containerised training jobs, and workflow orchestration. Labelling guidelines and inter-annotator agreement checks are part of the pipeline.
Optimisation
Bayesian and evolutionary hyperparameter search, and multi-objective tuning that trades accuracy against latency and model size when the target is an edge device.
Industry use cases
Typical applications by industry
These are representative applications of this service. They describe the kind of work we take on, not past clients.
Manufacturing
Surface-defect detection from line-scan cameras, and predictive maintenance from vibration and current signatures on rotating equipment.
Healthcare
Decision-support models for medical images and signals, validated on data from a separate site, with de-identification built into the pipeline. These tools assist clinicians and do not replace diagnosis.
Agriculture
Leaf-disease classification from phone and drone imagery that holds up across crop varieties and lighting, plus yield and irrigation models from weather and soil data.
Telecom
Traffic anomaly detection, capacity forecasting and fault prediction from network telemetry.
Government and universities
Multilingual document processing, analytics on public datasets, and reproducible research pipelines for funded projects.
Engagement models
Four ways to work with us
Start at the level that matches your certainty about the problem.
Consulting
A data and feasibility review, or an independent audit of an existing model or a vendor proposal, delivered as a written recommendation with risks.
Proof of concept
A time-boxed experiment on your data against an agreed baseline and metric. You receive code, a report and a recommendation to proceed or stop.
Full build
A production pipeline, the model, an API or edge deployment, tests, monitoring and documentation.
Ongoing support
Drift monitoring, retraining, re-labelling campaigns, dependency patches and new features.
Delivery process
How a AI & Machine Learning project runs
Frame and audit
Agree on the decision the model supports, the metric and the acceptable error. Audit data quality and independence.
Baseline and experiments
Build the simplest credible baseline, then test candidate models under the written protocol.
Engineer
Turn the selected model into a tested service or firmware component with input validation and logging.
Release
Run in shadow mode next to the current process, then release in stages with a rollback path.
Monitor and retrain
Track input drift, output drift and business metrics, and retrain from versioned data when thresholds are crossed.
FAQ
AI & Machine Learning questions
How much data do we need to start a machine learning project?
It depends on the task. With transfer learning, a few hundred labelled examples per class can be enough for a classification pilot. Detection and segmentation usually need more, and rare defects need a collection plan. During a proof of concept we plot learning curves, which show how much more data would actually help.
Can you work with data that cannot leave our premises?
Yes. Training can run on your own servers, we can work through remote access into your environment, and open-weight models remove the need to send text or images to an outside API. The access arrangement is agreed during discovery.
How do you make sure the reported accuracy holds in the field?
Test data is split by site, subject or time so that it is independent of training data. We check for duplicate and near-duplicate samples, report confidence intervals, and where possible test on data from a location the model has never seen. A shadow deployment then compares predictions with real outcomes before anyone relies on them.
Do you train large language models from scratch?
Rarely, because it is seldom justified. We start with prompting and retrieval over your documents using a pre-trained model, measure the result on an evaluation set, and fine-tune only when the numbers show a gap that fine-tuning can close.
Who maintains the model after release?
Either your team or ours. Every build includes drift monitors, alert thresholds and a retraining runbook. After the handover session your engineers can run it, or you can use our ongoing support model.
Can you review a model that another team built?
Yes. An audit covers data leakage, the evaluation protocol, calibration, subgroup performance, security exposure and reproducibility. You receive a report that separates defects from improvements that are optional.
Related services
Work that often goes with AI & Machine Learning
Cybersecurity Solutions
Threat modelling, security testing, secure architecture and detection engineering for IT, cloud, IoT and industrial control networks.
Explore CybersecurityAutonomous Robotics & Intelligent Systems
Perception, localisation, planning and control software for mobile robots, manipulators and drones, tested in simulation before hardware.
Explore Robotics & Intelligent SystemsFurther reading: Securing machine learning models in production
Have a dataset and a question?
Send a short description of the data and the decision you want to support. We will reply with what we would need to assess feasibility.